llm-tool-parser

A TypeScript utility for extracting and normalizing tool calls from large language model outputs.

It supports both native tool calling formats (OpenAI tool_calls, Claude tool_use, etc.) and “text-based tool calls” commonly produced by open-source or fine-tuned models.

The library converts inconsistent, streaming, or partially corrupted LLM outputs into a clean and unified ToolCall structure, making it easier to build reliable agent runtimes.

Key features

• Supports multiple LLM ecosystems (OpenAI, Claude, DeepSeek, Qwen, etc.)
• Extracts tool calls from raw text, JSON, XML, and hybrid formats
• Handles streaming, trailing text, and malformed outputs
• Normalizes all inputs into a unified ToolCall schema
• Works with both structured tool calling and prompt-based agents

Live parse

Live parse: https://saber2pr.top/llm-tool-parser/

Simple usage

Examples below are based on src/parser.test.ts.

import { parseLlmToolCalls } from 'llm-tool-parser'

const text =
  '{"tool_calls":[{"tool_name":"fileRead","arguments":{"filePath":"/tmp/a.ts"}}]}'

const { content, toolCalls } = parseLlmToolCalls(text)

console.log(content)
console.log(toolCalls[0]?.name)
console.log(toolCalls[0]?.arguments)

You can also parse model outputs that contain extra text:

const text =
  '让我来查看一下\n{"tool_calls":[{"tool_name":"fileRead","arguments":{"filePath":"/a.ts"}}]}'

const { content, toolCalls } = parseLlmToolCalls(text)

console.log(content) // 让我来查看一下
console.log(toolCalls)
/**
[
  {
    id: 'call_1780567643675_6a8b',
    name: 'fileRead',
    arguments: { filePath: '/a.ts' }
  }
]
 */

And code-fenced JSON is supported too:

const text =
  '我来读取文件\n```json\n{"tool_calls":[{"tool_name":"fileRead","arguments":{"filePath":"/a.ts"}}]}\n```'

const { content, toolCalls } = parseLlmToolCalls(text)

Supported parsable formats

Based on all test cases in src/parser.test.ts, the parser currently supports these input patterns:

1. OpenAI-style tool_calls

{"tool_calls":[{"tool_name":"fileRead","arguments":{"filePath":"/tmp/a.ts"}}]}

• single or multiple calls in one tool_calls array
• multiple consecutive JSON blocks
• JSON at the start with trailing text
• plain text before the JSON block
• JSON inside fenced code blocks

2. Single tool_name object

{"tool_name":"bash","arguments":{"command":"ls"}}

3. [tool_calls] marker + JSON objects

[tool_calls]{"tool":"grep","args":{"pattern":"hello"}}{"tool":"glob","args":{"pattern":"*.ts"}}

4. Action / Arguments text format

Action: fileRead
Arguments: {"filePath":"/tmp/test.ts"}

Also supports non-JSON arguments:

Action: grep
Arguments: foo

5. XML-style tool tags

<fileRead>{"filePath":"/tmp/a.ts"}</fileRead>

• ignores non-tool tags like <thinking>
• supports XML inside fenced code blocks

6. CLI log style [Called tools: ...]

● [Called tools: fileRead({"filePath":"/tmp/a.ts"})]

Also supports multiple calls:

● [Called tools: grep({"pattern":"foo"}), glob({"pattern":"*.ts"})]

7. OpenAI legacy function_call

{
  "function_call": {
    "name": "fileRead",
    "arguments": "{\"filePath\":\"/tmp/a.ts\"}"
  }
}

8. Claude tool_use

Content array form:

{
  "content": [
    {
      "type": "tool_use",
      "name": "fileRead",
      "input": {"filePath": "/tmp/a.ts"}
    }
  ]
}

Standalone form:

{
  "type": "tool_use",
  "name": "fileRead",
  "input": {"filePath": "/tmp/a.ts"}
}

9. OpenAI Responses API function_call

{
  "type": "function_call",
  "name": "fileRead",
  "arguments": "{\"filePath\":\"/tmp/a.ts\"}"
}

10. Simple tool + args object

{"tool":"fileRead","args":{"filePath":"/tmp/a.ts"}}

11. Root array of tool calls

[
  {"tool":"grep","args":{"pattern":"foo"}},
  {"tool":"glob","args":{"pattern":"*.ts"}}
]